minimum residual based model order reduction approach for

Minimum residual based model order reduction approach for

unsteady nonlinear aerodynamic problems M. Ripepi1, R. Zimmermann2 and S. Görtz1

1German Aerospace Center (DLR), Institute of Aerodynamics and Flow Technology, Braunschweig, Germany 2Institute Computational Mathematics, TU Braunschweig, Braunschweig, Germany

Data-driven Model Order Reduction and Machine Learning Workshop University of Stuttgart, Germany 30 March 2016 – 1 April 2016

1. Motivation and objectives

2. Model order reduction for unsteady CFD

3. Accelerated greedy MPE procedure

4. Aeronautical application: the LANN wing

5. Summary and outlook

DLR.de • Chart 2

Overview

Model order reduction for unsteady aerodynamic problems > Matteo Ripepi • MORML 2016 workshop > University of Stuttgart, Germany. 30 Mar.–1 Apr. 2016

Prediction of: Aerodynamic performance "Handling Qualities" Structural loads

Input for: Aerodynamic design Structural design and sizing Design of control surfaces

and flight control system Flight simulators

flaps up

stall area

stall area

Negative limit load factor

positive limit load factor

Cruise Velocity

Motivation From design to certification of an aircraft many aerodynamic data are

needed – for the entire flight envelope – All required aerodynamic data can be derived from the pressure and

shear stress distribution on the aircraft surface, from steady and unsteady simulations

DLR.de • Chart 3

Compressibility

Unsteady effects

Separated flows

normal operation range

CFD is mostly limited to the cruise point


CFD too expensive to procure all the necessary Aero Data (in particular in multidisciplinary settings requiring repeated simulations)

Objectives

Development of a simulation environment, where ALL the necessary aerodynamic data can be determined numerically

with high accuracy with Hi-Fi methods

Result!

Solution(𝑝𝑝1, 𝑝𝑝2, …)?

DLR.de • Chart 4

ROMs for compressible flows

To reduce the computational complexity, Reduced Order Models based on Hi-Fi CFD should replace (e.g. in the MultiDisciplinary Optimization)

high quality models and methods

ROM


CFD

Objectives

Development of a simulation environment, where ALL the necessary aerodynamic data can be determined numerically

with high accuracy with Hi-Fi methods

DLR.de • Chart 5

ROMs for compressible flows


Reduced Order Model

Snapshots 𝑆𝑆𝑖𝑖(𝑝𝑝𝑖𝑖1, … )

Prediction 𝑉𝑉�(𝑝𝑝1, … )

𝛼𝛼 = 0°

𝛼𝛼 = 2°

𝛼𝛼 = 6°

𝛼𝛼 = 4° Offline

Online

Restricted range of validity

Low-dim. (r << n ODEs) description of large-scale

system dynamics

Wide range of validity

Full-order model, high-fidelity CFD data (system of n ODEs)

ROMs for the prediction of steady aerodynamic flows

Model order reduction approach for unsteady aerodynamic applications

Model order reduction for unsteady CFD

𝐰𝐰 t10 𝐰𝐰 t60 𝐰𝐰 t110 ⋯ ⋯

Search for an approximated solution 𝐰𝐰 = [ρ, ρ𝐯𝐯, ρ𝐸𝐸𝑡𝑡] ∈ ℝ𝑁𝑁 in the POD subspace 𝐔𝐔 ∈ ℝ𝑁𝑁 x 𝑟𝑟 , 𝑟𝑟 ≪ 𝑁𝑁 minimizing the unsteady residual in the L2 norm

𝐰𝐰 ≈�𝑎𝑎𝑖𝑖𝐔𝐔𝑖𝑖

𝑟𝑟

𝑖𝑖=1

+ 𝐰𝐰 = 𝐔𝐔𝐔𝐔 + 𝐰𝐰

𝐰𝐰 =∑ 𝐰𝐰 ti𝑚𝑚𝑖𝑖=1𝑚𝑚

: average of the snapshots

𝐔𝐔: unknown coefficients

DLR.de • Chart 6 Model order reduction for unsteady aerodynamic problems > Matteo Ripepi • MORML 2016 workshop > University of Stuttgart, Germany. 30 Mar.–1 Apr. 2016

N: order of CFD model (variables x nodes)

r: order of the ROM (i.e. number of POD modes)

Proper Orthogonal Decomposition

𝐘𝐘 = 𝐰𝐰 t1 , . . . ,𝐰𝐰 tm −𝐰𝐰

SVD: 𝐘𝐘 = 𝐔𝐔𝐒𝐒𝐕𝐕𝐓𝐓 EVD: 𝐑𝐑 = 𝐘𝐘𝐘𝐘𝐓𝐓 = 𝐔𝐔𝐒𝐒𝟐𝟐𝐔𝐔𝐓𝐓

min𝐔𝐔

∥ 𝐑𝐑� 𝐔𝐔𝐔𝐔 + 𝐰𝐰 ∥L22

𝐑𝐑� ≝ 𝐑𝐑 𝐰𝐰 t + 𝛀𝛀𝜕𝜕𝐰𝐰 t𝜕𝜕𝜕

= 0 ∈ ℝ𝑁𝑁 𝛀𝛀: cell volumes Semi-discrete Navier-Stokes Eqs.

𝐑𝐑� ≝ 𝐑𝐑 𝐰𝐰rom tn+1 + 𝛀𝛀3𝐰𝐰rom tn+1 − 4𝐰𝐰rom tn + 𝐰𝐰rom tn−1

2Δ𝜕

Per each physical time step, find the POD coefficients minimizing the unsteady residual at the time step n+1, by solving a nonlinear least squares problem: 𝐰𝐰rom tn+1 ≈ 𝐔𝐔𝐔𝐔+ 𝐰𝐰

𝐉𝐉T𝐉𝐉 + 𝛌𝛌𝛌𝛌 Δ𝐔𝐔 = −𝐉𝐉T𝐑𝐑�

DLR.de • Chart 7

Model order reduction for unsteady CFD

Broyden’s method for updating 𝐉𝐉

𝐉𝐉𝑘𝑘+1 = 𝐉𝐉𝑘𝑘 +∆𝐑𝐑� − 𝐉𝐉𝑘𝑘 Δ𝐔𝐔

Δ𝐔𝐔 2 Δ𝐔𝐔T

N: order of CFD model (variables x nodes)


Updating pseudo-Hessian matrix 𝐀𝐀 ≡ 𝐉𝐉T𝐉𝐉 ∈ ℝ𝒓𝒓×𝒓𝒓 ⇒ 𝒪𝒪 Nr2

𝐀𝐀𝑘𝑘+1 ≡ 𝐉𝐉𝑘𝑘+1T 𝐉𝐉𝑘𝑘+1

= 𝐀𝐀𝑘𝑘 + 𝐉𝐉𝑘𝑘T ∆𝐑𝐑� − 𝐀𝐀𝑘𝑘 Δ𝐔𝐔

Δ𝐔𝐔 2 Δ𝐔𝐔T + Δ𝐔𝐔∆𝐑𝐑�T 𝐉𝐉𝑘𝑘 − Δ𝐔𝐔T𝐀𝐀𝑘𝑘

Δ𝐔𝐔 2

+ Δ𝐔𝐔∆𝐑𝐑�T∆𝐑𝐑� − ∆𝐑𝐑�T𝐉𝐉𝑘𝑘 Δ𝐔𝐔 − Δ𝐔𝐔T𝐉𝐉𝑘𝑘T ∆𝐑𝐑� + Δ𝐔𝐔T𝐀𝐀𝑘𝑘Δ𝐔𝐔

Δ𝐔𝐔 4 Δ𝐔𝐔T

⇒ 𝒪𝒪 Nr

𝐉𝐉 ∈ ℝ𝑵𝑵×𝒓𝒓


𝐔𝐔 ← 𝐔𝐔 + Δ𝐔𝐔

until

co

nver

genc

e

Training maneuvers Schroeder multisine with 0.01 < k < 0.2

16430 elements 21454 points

RANS equations with Spalart-Allmaras turbulence model Mach = 0.8, Re = 7.5 106

Training maneuvers exciting up to k ≝ 𝜔𝜔𝜔𝜔𝑉𝑉∞

= 0.2

Prediction for periodic motions at different frequencies and amplitudes (using the same ROM)

Linear chirp signal with kmax = 0.2

NACA 64A010

DLR.de • Chart 8

2D Test Case: NACA airfoil Model order reduction for unsteady aerodynamic problems > Matteo Ripepi • MORML 2016 workshop > University of Stuttgart, Germany. 30 Mar.–1 Apr. 2016

2D Test Case: NACA airfoil Periodic pitching motion prediction at different oscillation frequencies


ROM built with 50 POD modes

0.08

0.24 k

0.08

0.24

k


Online Performances TAU ROM

Chirp Schroeder

Average CPU Time

total 15 min 58 s 8 min 58 s 8 min 43 s per time step 6.39 s 3.56 s 3.49 s speed-up 1 1.79 1.83

ROM built with 50 POD modes

2D Test Case: NACA airfoil Periodic pitching motion prediction at different oscillation amplitudes

2°

3°

αA αA

3°

2°

Ways to improve the performances:

Hyper-reduction (e.g. Gappy POD) Different selection of the subspace

(e.g. Dynamic Mode Decomposition)

I. compute the approx. solution

II. evaluate the unsteady residual

III. solve the LS problem


Model order reduction for unsteady CFD Curse of dimensionality in ROMs of nonlinear systems

𝐑𝐑� 𝐰𝐰 t ≝ 𝐑𝐑 𝐰𝐰 t + 𝛀𝛀𝜕𝜕𝐰𝐰 t𝜕𝜕𝜕

𝐰𝐰 t ≈ 𝐔𝐔 𝐔𝐔 t + 𝐰𝐰 ⇒ 𝒪𝒪 Nr N: order of CFD model (variables x nodes)


Jacobian Matrix 𝐉𝐉 ∈ ℝ𝑵𝑵×𝒓𝒓

nonlinear least squares problem 𝑚𝑚𝑚𝑚𝑚𝑚𝐔𝐔

∥ 𝐑𝐑� 𝐔𝐔𝐔𝐔 + 𝐰𝐰 ∥L22

The computational cost scales linearly with the dimension N of the full order model. No significant speedup can be expected when solving the minimum residual ROM.

𝐉𝐉T𝐉𝐉 + 𝛌𝛌𝛌𝛌 Δ𝐔𝐔 = −𝐉𝐉T𝐑𝐑� ⇒ 𝒪𝒪 Nr

i. Compute RHS −𝐉𝐉T𝐑𝐑� ii. Broyden’s method for updating 𝐉𝐉 iii. Updating pseudo-Hessian matrix

𝐀𝐀 ≡ 𝐉𝐉T𝐉𝐉 ∈ ℝ𝒓𝒓×𝒓𝒓


Hyper-reduction approaches

Selecting the subset indices Gappy POD Missing Point Estimation (Discrete) Empirical Interpolation Method

Complexity reduction by sampling (or compute only a few entries of) the nonlinear unsteady residual vector 𝐑𝐑�

𝐑𝐑�

𝐑𝐑��

𝐑𝐑�∗

Reconstruction approximation of the entire

vector, by interpolation or by least-squares projection onto a subspace 𝐕𝐕 = U 𝐔𝐔𝑇𝑇P 𝐏𝐏𝑇𝑇U −1𝐔𝐔𝑇𝑇P

Collocation omission of many components non intrusive

= V

𝐏𝐏

mask matrix

𝐏𝐏T𝐑𝐑�

U 𝐔𝐔𝑇𝑇P 𝐏𝐏𝑇𝑇U −1𝐔𝐔𝑇𝑇P 𝐏𝐏𝑇𝑇= 1/𝜎𝜎min 𝐏𝐏𝑇𝑇𝐔𝐔

Greedy: minimize

The complete nonlinear unsteady residual vector 𝐑𝐑� is evaluated, but only a small subset of its entries are used in the minimization process

Exhaustive greedy MPE maximize 𝜎𝜎min 𝐏𝐏𝑠𝑠+1𝑇𝑇 𝐔𝐔 Input: 𝐔𝐔 ∈ ℝ𝑁𝑁×𝑟𝑟: basis of a r-dim subspace,

𝐉𝐉𝑠𝑠 ∈ ℝ𝑠𝑠×1: index set, 𝐏𝐏𝑠𝑠 ∈ ℝ𝑁𝑁×𝑠𝑠: mask matrix with s indices, where s ≥ r.

Output: next index 𝐉𝐉𝑠𝑠+1 = 𝐉𝐉𝑠𝑠 ∪ 𝑗𝑗opt , 𝐏𝐏𝑠𝑠+1𝑇𝑇 = 𝐏𝐏𝑠𝑠, 𝐞𝐞𝑗𝑗, opt


Exhaustive greedy missing point estimation procedure

𝐉𝐉𝑠𝑠 2 5 7

𝐏𝐏𝑠𝑠

𝐞𝐞2 𝐞𝐞5 𝐞𝐞7

Greedy point selection algorithms minimize an error indicator by sequentially looping over all entries costly ⇒ > 𝒪𝒪 Nr3

Singular Value Decomposition

𝐯𝐯𝑗𝑗 ≔ 𝐞𝐞𝑗𝑗, 𝑠𝑠+1𝑇𝑇 𝐔𝐔𝚽𝚽𝑠𝑠

𝑇𝑇

Rank-one SVD update

𝐔𝐔𝑇𝑇𝐏𝐏𝑠𝑠+1𝐏𝐏𝑠𝑠+1𝑇𝑇 𝐔𝐔 = 𝚽𝚽𝒔𝒔 𝚺𝚺𝑠𝑠2 + 𝐯𝐯𝑗𝑗𝐯𝐯𝑗𝑗𝑇𝑇 𝚽𝚽𝑠𝑠𝑇𝑇

0 < 𝜇𝜇𝑖𝑖 < 1 ∑ 𝜇𝜇𝑖𝑖𝑟𝑟

𝑖𝑖=1 = 1 𝜆𝜆𝑖𝑖 = 𝜎𝜎𝑖𝑖2 + 𝜇𝜇𝑖𝑖 𝐯𝐯 2

Eigenvalues

Symmetric rank-one eigenvalue modification

R. Zimmermann and K. Willcox. An accelerated greedy missing point estimation procedure. SIAM Journal on Scientific Computing, 2/2016. Preprint.

The penultimate singular value bounds the growth of 𝜎𝜎min

Observation 2)

Observation 1)


Accelerated greedy missing point estimation procedure

R. Zimmermann and K. Willcox. An accelerated greedy missing point estimation procedure. SIAM Journal on Scientific Computing, 2/2016. Preprint.

Select the vector 𝐯𝐯𝑗𝑗 leading to the largest growth in the smallest eigenvalue 𝜆𝜆min 𝐯𝐯𝑗𝑗 of 𝐌𝐌 ≔ 𝚺𝚺𝑠𝑠2 + 𝐯𝐯𝑗𝑗𝐯𝐯𝑗𝑗𝑇𝑇 ∈ ℝ𝑟𝑟×𝑟𝑟

Build fast approximations that sort the set of candidate vectors that induce the rank-one modifications ( without solving the modified eigenvalue pb.)

Fast greedy MPE with target switching based on the growth potential ⇒ 𝒪𝒪 Nr2 Input: 𝐔𝐔 ∈ ℝ𝑁𝑁×𝑟𝑟: basis of a r-dim subspace, 𝐉𝐉𝑠𝑠 ∈ ℝ𝑠𝑠×1: index set, s ≥ r.

1. �̅�𝐉s = 1, … ,𝑁𝑁 \ 𝐉𝐉𝑠𝑠 2. while 𝐉𝐉𝑠𝑠 ≤ maxpoints do 3. target the largest growth for 𝜆𝜆min (i.e. 𝜆𝜆r) or the next biggest 𝜆𝜆r−1, etc. 4. determine 𝑗𝑗opt by sorting 𝐯𝐯𝑗𝑗 , 𝑗𝑗 ∈ �̅�𝐉s according to the fast estimate 5. update: 𝐉𝐉s+1 = 𝐉𝐉𝑠𝑠 ∪ 𝑗𝑗opt , �̅�𝐉s+1 = �̅�𝐉s \ 𝑗𝑗opt , 𝑠𝑠 = 𝑠𝑠 + 1 Output: index set 𝐉𝐉𝑠𝑠

Idea:

𝜆𝜆r−𝑘𝑘 < ℱ(𝜎𝜎𝑟𝑟−𝑘𝑘2 ,𝜎𝜎𝑟𝑟−𝑘𝑘+12 , 𝑣𝑣𝑖𝑖) Error bound

Select the vectors 𝐯𝐯𝑗𝑗, opt that feature the largest absolute values in the ultimate component while all other components are comparably small.


500 MPE nodes 50 POD modes Total number of grid nodes: 21454

2D Test Case: NACA airfoil Nodes selected with the fast greedy MPE

2D Test Case: NACA airfoil Periodic pitching motion prediction at k=0.16


Maximum lift minimum lift

1000 nodes selected by the

greedy MPE

< 5% of the total number of grid nodes


Online Performances (1 core)

CFD TAU

ROM (built with 50 POD modes)

w/o MPE with MPE

100 nodes 500 nodes 1000 nodes

CPU Time

Total [min] ~18 ~9 ~5 ~5 ~6 per time step [s] 6.98 3.65 2.10 2.17 2.30

speed-up 1 1.91 3.31 (+42%)

3.22 (+41%)

3.03 (+37%)

Offline Performances (1 core) Compute 400 snapshots by a training unsteady CFD simulation: ~47 min POD ROM building: 1min 32s Fast greedy MPE (selecting 1000 nodes): ~2 min

2D Test Case: NACA airfoil Performances

Process CFD TAU

ROM (built with 50 POD modes) w/o MPE with MPE (1000 nodes)

CPU Time

Building the model — ~49 min ~51 min Simulate 7 parameters (5 reduced frequencies and 2 amplitudes)

2 h 5 min 1 h 3 min 42 min

450560 elements 469213 points

3D test case: LANN wing

DLR.de • Chart 18

RANS equations with SA-neg turbulence model V∞ = 271.66 m/s, Mach = 0.82, Re = 7.17 106

Chirp training maneuver exciting up to k ≝ 𝜔𝜔𝜔𝜔𝑟𝑟𝑉𝑉∞

= 0.35

Predicted periodic pitching oscillation: 𝛼𝛼 𝜏𝜏 = 2.6° + 0.25 sin(𝐤𝐤 𝜏𝜏) , τ ≝ 𝑉𝑉∞𝑡𝑡𝜔𝜔𝑟𝑟


CFD settings Min residual: 1e-5 Max inner iterations: 100

ROM settings 18 POD modes 5000 MPE nodes


Total number of grid nodes: 469213

5000 nodes selected by the greedy MPE


~1% of the total number of grid nodes

k=0.3



Online Performances CFD

(10 cores)

ROM (10 cores)

w/o MPE with MPE

Average Wall Time

total 25 h 50 min 2 h 44 min 1 h 29 min per time step ~16 min 1 min 36 s 36 s speed-up 1 9.5 26.7

5000

Compute snapshots: 4d 10h 40min POD ROM building: ~5 min Fast greedy MPE: ~6 min

Offline Performances (10 cores)

The effectiveness of reduced order models for unsteady aerodynamic problems have been demonstrated for a NACA airfoil and the LANN wing in transonic flows.

Remarks

Training maneuvers may be an issue: the quality of the resulting model will reflect the input signal quality

Further analyses are required (e.g. moving window, full-aircraft config., …)

Summary and Outlook

Source: DLR-AS

MDO: 2.5g maneuver with aileron deflection

Assessment of Spoiler Concepts

DLR.de • Chart 21

Aeroelastic Gust Loads

Possible improvements Different selection of the subspace (e.g. Dynamic Mode

Decomposition, Isomap) Different ordering of the modes, focusing more on the

input-output behavior (e.g. subspace rotation) Hybrid strategy by switching in time ROM/CFD Sparsity promoting techniques (L1 norm, clustering)


Scenarios

Multidisciplinary applications

Real-time applications

Design and optimization

Probabilistic applications

ROMs


Acknowledgement to the EU: Part of the research leading to this work was supported by the AEROGUSTproject, funded by the European Commission under grant number 636053. COST Action TD1307: EU-MORNET, Model Reduction Network.

German National Research Projects Lufo-IV joint research project AeroStruct

Internal DLR Projects DLR multidisciplinary project Digital-X

Acknowledgements


Bibliography

DLR.de • Chart 24

T. Franz, R. Zimmermann, S. Görtz, and N. Karcher. Interpolation-based reduced-order modelling for steady transonic flows via manifold learning. InternationalJournal of Computational Fluid Dynamics, 28(3–4):106–121, 2014.

A. Vendl and H. Faßbender. Model order reduction for unsteady aerodynamicapplications. In Proc. Appl. Math. Mech., volume 13, pages 475–476, 2013.

A. Vendl, H. Faßbender, S. Görtz, R. Zimmermann, and M. Mifsud. Model orderreduction for steady aerodynamics of high-lift configurations. CEAS AeronauticalJournal, 5(4):487–500, 2014.

R. Zimmermann and S. Görtz. Non linear reduced order models for steadyaerodynamics. In Procedia Computer Science, volume 1, pages 165–174, 2010.Proceedings of the International Conference on Computational Science, ICCS 2010.

R. Zimmermann and K. Willcox. An accelerated greedy missing point estimationprocedure. SIAM Journal on Scientific Computing, 2/2016. Preprint.


minimum residual based model order reduction approach for

Documents